A method, system and device for super-resolution reconstruction of magnetic resonance images

Abstract

The present invention provides a method, system and device for super-resolution reconstruction of magnetic resonance images, mainly including the following steps: segmenting a 3D magnetic resonance image into multiple 2D magnetic resonance images; using the 2D magnetic resonance images as single-channel inputs and deeply extracting their features by means of a neural network; performing upsampling on the features using an upsampling structure with shared parameters; and outputting a high-resolution magnetic resonance image after integrated processing. The combination of the cascade module and multi-scale upsampling in this solution can extract high- and low-frequency features of different depths, enabling the model to have sufficient expressive power, and sharing the upsampling structure parameters to greatly lighten the model. Experiments prove that the method of the present invention is superior to the prior art in terms of reconstruction quality, imaging time and model parameters, reduces the examination time, and can generate image details consistent with human vision, avoiding artifacts that are not conducive to image observation.

Description

Technical Field

[0001] The present invention relates to the technical field of image enhancement, and in particular, to a method, system and device for super-resolution reconstruction of magnetic resonance images. Background Art

[0002] Currently, magnetic resonance imaging (MRI) is widely used in the field of medical imaging. Compared with computed tomography (CT / PET), MRI has no radiation damage and can be used for multi-faceted and multi-parameter imaging. However, due to factors such as hardware equipment, scanning time, signal-to-noise ratio, and patient body movement, it is often difficult to quickly obtain high-quality MRI. At the same time, long-term scanning may cause discomfort to some patients.

[0003] In a clinical study conducted by New York University in cooperation with Facebook, a team of computer scientists and radiologists demonstrated that deep learning methods can generate MRI with the same high precision using approximately one-fourth of the original data required for traditional complete MRI. Since less data is needed, the operating speed of the magnetic resonance machine scan is almost 4 times faster. The super-resolution method is expected to improve the quality of magnetic resonance images without upgrading the hardware.

[0004] In previous studies, traditional super-resolution algorithms were generally divided into two categories: interpolation-based algorithms and learning-based algorithms. Interpolation algorithms have been widely used in the super-resolution reconstruction of medical images. However, since the pixels of high-definition images are interpolated from the pixels of low-definition images according to a specific mathematical method, they are often blurred at the details. Learning-based methods usually require a large amount of data as support. If the data is incomplete or mismatched (such as using natural images as a dictionary to represent MR images), the image effect is often not good.

[0005] With the development of artificial intelligence, various deep learning-based algorithms have become research hotspots in the field of super-resolution. Among them, methods based on convolutional neural networks (CNNs) include SRCNN and FSRCNN, which were proposed and applied to the super-resolution problem of MR images earlier. Algorithms VDSR and EDSR adopt a residual structure to prevent the network from losing low-frequency information and construct deeper networks. On this basis, algorithms RCAN and SAN introduce a channel attention mechanism and a second-order attention mechanism to improve network efficiency and achieve good results in natural images. To quickly reconstruct images, many methods have also considered the lightweight of the network, such as CARN, LESRCNN, IMDN, and RFDN. These methods focus on efficiently completing the super-resolution task with fewer parameters. Generative adversarial networks (GANs) have also done a lot of work in the field of super-resolution and have been applied to the super-resolution task of MRI. By introducing adversarial loss, perceptual loss, and ranking loss, the image details are made more abundant and more in line with the human vision.

[0006] However, the main purpose of these methods is to improve the performance metrics of the natural image super-resolution task, and there is no targeted improvement for the MR image super-resolution task. For the MR image super-resolution task, it should be considered to shorten the imaging time while generating high-resolution images, reduce the pain of patients during the examination, and generate image details that are as consistent with human vision as possible, and avoid generating artifacts that are not conducive to doctors' observation. Although generative adversarial networks have proven in many experiments that the reconstructed images are more visually consistent with human perception, the network structure is usually large. For lightweight networks, due to the limitation of the number of parameters, they usually do not have sufficient expressive power, and the adversarial loss will affect the main metrics (PSNR and SSIM) of the images. For deeper residual networks, the next step can only be carried out after the calculation at the previous node is completed, which requires a long imaging time and results in a long waiting time. Summary of the Invention

[0007] The purpose of the present invention is to provide a method, system and device for super-resolution reconstruction of magnetic resonance images to solve at least one of the above technical problems existing in the prior art.

[0008] To solve the above technical problems, a method for super-resolution reconstruction of magnetic resonance images provided by the present invention includes the following steps:

[0009] Step 1: Segment the 3D MR image into 2D MR images as the single-channel input of the super-resolution network.

[0010] Preferably, during network training, the high-definition MR image is downsampled 2, 4, or 8 times by bicubic interpolation as the input, so as to create smoother image edges.

[0011] Preferably, during network training, the high-definition MR image is downsampled 2, 4, or 8 times in the frequency domain as the input, so as to facilitate analysis and processing.

[0012] In addition to the above algorithms and parameters, other well-known algorithms and related parameters in the art can also be used for training.

[0013] Step 2: Based on the single-channel input, convert it into a multi-channel feature map I LR , and using multiple cascaded blocks and convolutional blocks arranged alternately, obtain the feature maps output by convolutional blocks with different depths and the feature maps output by the cascaded blocks The feature relationship of can be expressed recursively as where B i represents the i-th convolutional block, C i represents the i-th cascaded block, and x represents the input image.

[0014] Preferably, the inside of the cascade block is a cascade structure composed of 3 residual blocks and 3 convolutional blocks, and its internal features are still consistent with the recursive form of the external global cascade structure.

[0015] Preferably, in the above operation, the value range of i is a positive integer where i ≤ 8, that is, the upper limit number of convolutional blocks and cascade blocks that can be set is 8.

[0016] In addition to the above algorithms and parameters, other well-known algorithms and related parameters in the art can also be used for feature extraction.

[0017] Step 3: Weight the spatial positions of the feature maps output by each convolutional block to obtain the processed features

[0018] Preferably, a spatial attention mechanism is used to weight the spatial positions of the feature maps output by each convolutional block, and max pooling MaxPool and average pooling Avgpool are used to process the features at the same spatial position of each channel in, fuse the feature maps and then perform convolution processing, and then convert the obtained features into weights between 0 and 1 by the sigmoid function.

[0019] Preferably, the above operation can adopt a spatial attention mechanism with different or not completely the same parameters.

[0020] Preferably, the convolutional block in the spatial attention mechanism adopts a 7x7 convolutional block.

[0021] The above process can be expressed as:

[0022]

[0023] Use a sub-pixel convolutional block with parameter sharing to perform multi-scale upsampling on the feature maps with different depths processed by the spatial attention mechanism to obtain multiple groups of super-resolution feature maps

[0024] Preferably, the above operation can adopt a non-shared or not completely shared sub-pixel block or transposed convolutional block.

[0025] In addition to the above algorithms and parameters, other well-known algorithms and related parameters in the art can also be used for sampling.

[0026] Step 4: Use the information integration block to perform integration processing on the multiple feature maps obtained by multi-scale upsampling and output a single-channel ultra-clear MR image I SR 。

[0027] Preferably, the information integration block is linearly and alternately composed of 3, 4, or 5 groups of activation functions RELU and convolutional blocks such as 3×3.

[0028] Preferably, one or more features with different depths can be selected for integrated processing in the above operation.

[0029] In addition to the above algorithms and parameters, other well-known algorithms and related parameters in the art can also be used for information integration.

[0030] On the other hand, the present invention also discloses a magnetic resonance image super-resolution reconstruction system, including an image receiving module, a feature extraction module, a sampling module, an information integration module, and an image generation module:

[0031] The image receiving module receives 3D magnetic resonance images and segments them into 2D images, which are sent to the feature extraction module as a single channel.

[0032] The feature extraction module receives the 2D images, performs depth feature extraction, outputs feature maps, and sends them to the sampling module.

[0033] The sampling module receives the feature maps, weights their spatial positions using a spatial attention mechanism, and performs multi-scale upsampling to obtain multiple groups of super-resolution feature maps, which are sent to the information integration module.

[0034] The information integration module performs integrated processing on the super-resolution feature maps and outputs a single-channel ultra-clear image to the image generation module.

[0035] The image generation module outputs the ultra-clear image.

[0036] On the other hand, the present invention also provides a magnetic resonance image super-resolution reconstruction device, mainly including a processor, a memory, and a bus. The memory stores instructions that can be read by the processor, and the processor is used to call the instructions in the memory to execute a magnetic resonance image super-resolution reconstruction method. The bus connects the functional components to transmit information.

[0037] Adopting the above technical solutions, the present invention has the following beneficial effects:

[0038] A magnetic resonance image super-resolution reconstruction method, system, and device provided by the present invention adopt the super-resolution reconstruction technology in deep learning to achieve accurate and clear super-resolution reconstruction of magnetic resonance images while realizing model lightweight and fast imaging. The method of the present invention constructs a global cascade structure to extract features of low-resolution magnetic resonance images. The input of each convolutional block includes all the features output by the previous cascade blocks, which ensures that the features output by each convolutional block have a large degree of similarity.

[0039] By using the same spatial attention structure to process the features output by convolutional blocks with different depths, and then performing multi-scale upsampling on these features with different depths using a sub-pixel convolutional block with parameter sharing, the model has sufficient expressive ability while being lightweight.

[0040] Through experimental verification, the method proposed by the present invention has obvious advantages in terms of the accuracy (PSNR / SSIM), speed, and parameter memory of generating high-definition magnetic resonance images compared with other existing technologies (such as SRGAN, CARN, SAN, RFDN, LESRCNN), reducing the examination time of patients, alleviating the examination pain, generating image details that conform to human vision, and avoiding the generation of artifacts that are not conducive to observation. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] In order to more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the following will briefly introduce the drawings required for the description of the specific embodiments or the prior art. Obviously, the drawings in the following description are some embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.

[0042] Figure 1 It is a structural diagram of an embodiment of the present invention;

[0043] Figure 2 It is a detailed structural diagram of the designed network of the present invention;

[0044] Figure 3 It is a comparison diagram of the effects of the present invention and the prior art;

[0045] Figure 4 It is a system diagram of a magnetic resonance image super-resolution reconstruction. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0046] The following will clearly and completely describe the technical solutions of the present invention with reference to the drawings. Obviously, the described embodiments are some, but not all, of the embodiments of the present invention. All other embodiments obtained by those of ordinary skill in the art without creative efforts based on the embodiments of the present invention belong to the scope of protection of the present invention.

[0047] In the description of the present invention, it should be noted that the terms "center", "upper", "lower", "left", "right", "vertical", "horizontal", "inner", "outer", etc. indicate the orientation or positional relationship based on the orientation or positional relationship shown in the drawings. It is only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of the present invention. In addition, the terms "first", "second", "third" are only used for descriptive purposes and cannot be construed as indicating or implying relative importance.

[0048] In the description of the present invention, it should be noted that unless otherwise clearly specified and defined, the terms "installation", "connection", and "coupling" should be understood in a broad sense. For example, it can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be directly connected or indirectly connected through an intermediate medium, and it can be the communication inside two components. For those of ordinary skill in the art, the specific meanings of the above terms in the present invention can be understood according to specific situations.

[0049] The following further explains the present invention with specific embodiments.

[0050] The present invention proposes a lightweight multi-stage upsampling mechanism, which extracts features of an image through a cascaded structure, so that features at different depths always have similar parts to a certain extent and effectively reduces the depth of the network. Then, the spatial attention structure weights the same spatial positions, so that the high-frequency information that should be restored more is concerned. After that, an upsampling structure with parameter sharing is used to perform multi-scale upsampling on features at different depths, so that the network has sufficient expressive power under the conditions of low parameter quantity and low depth. Compared with the existing networks currently, the method proposed by us has obvious advantages in terms of parameter quantity, running time, and image quality (PSNR\SSIM).

[0051] Specifically, the MR image super-resolution reconstruction method of the present invention includes three steps: feature extraction, multi-scale upsampling, and information integration. In a preferred embodiment, feature extraction gradually deepens the extraction of the features of the input single-channel image through a globally cascaded convolutional block and a cascaded residual block. The selection of the convolutional block can be set based on the needs of deep feature extraction, output data dimension requirements, etc. For example, a 3×3 convolutional block, 5×5, etc. can be used. As the feature extraction deepens, the network will pay more and more attention to high-frequency information, and due to the cascaded structure, there are always similar parts in the features at different depths. Multi-scale upsampling includes two parts. First, the same weights are applied to the spatial positions of the features at different depths, so that similar high-frequency information is concerned. In a preferred embodiment, in this stage, the spatial attention mechanism provided in this embodiment can be used for weighting. Then, an upsampling module with parameter sharing (transpose convolution or sub-pixel convolution) is used to perform multi-scale upsampling on the features at different depths. Information integration is composed of multiple 3×3 convolutional blocks and the activation function RELU, and integrates the features obtained by multi-scale upsampling through a non-linear transformation.

[0052] The detailed steps of feature extraction, multi-scale upsampling, and information integration in the super-resolution reconstruction of MR images are as follows, as Figure 1 shown:

[0053] 1. Use the feature map output by the first 3×3 convolutional block as the input of the next cascading block, and then use the feature map output by the cascading block as the input of the next convolutional block. And so on, to obtain the features output by convolutional blocks with different depths and the features output by the cascading blocks In a preferred embodiment, except that the input of the first convolutional block is single-channel, the inputs and outputs of other filters are all 64 channels.

[0054] 2. Obtain the features output by convolutional blocks with different depths and the features output by the cascading blocks The feature relationship can be represented recursively as where B i represents the i-th convolutional block, C i represents the i-th cascading block, and x represents the input image. Inside the cascading block, the features are still passed in a cascading manner, consisting of a residual block and a 3×3 convolutional block, as Figure 2 shown. The inputs and outputs of all filters are 64 channels, and the residual block uses grouped convolution.

[0055] 3. Use a spatial attention structure, as Figure 2 shown, to perform the same weighting on the spatial positions of the features f CONV output by each 3×3 convolutional block. Use max pooling MaxPool and average pooling Avgpool to process the features at the same spatial positions of each channel in, fuse the feature maps, process them through a 7×7 convolution, and then convert the obtained features into weights between 0 and 1 by a sigmoid function. Finally, perform weighting on the input at the corresponding spatial positions to obtain the features after the attention mechanism processing The inputs and outputs of all filters are 64 channels.

[0056] Then use a sub-pixel block or transposed convolution with parameter sharing to upsample the features with different depths after the spatial attention mechanism processing at multiple scales to obtain multiple groups of super-resolution feature maps where the size of the feature maps is increased to 2, 4, or 8 times that of the input according to the setting, and the number of feature channels remains unchanged (both the input and output are 64).

[0057] This upsampling process takes into account the low-frequency and high-frequency features at different depths, prevents the loss of low-frequency features during the feature depth extraction process, and at the same time highlights the high-frequency features by the spatial attention mechanism.

[0058] 4. Use an information integration block for the multiple feature maps obtained by multi-scale upsampling Integrated processing, through non-linear transformation of multiple (3 to 5) alternately cascaded 3×3 convolutional blocks and the activation function RELU, finally outputs a single-channel ultra-clear MR image I by the convolutional block SR 。

[0059] In a more specific embodiment, we conduct a comparison of experimental effects in the following environment: the operating environment is windows, implemented in pytorch1.7, and the FASTMRI dataset and IXI dataset are selected for training and testing. The MRI acquisition protocols include: T1, T2, and PD weighted images, MRA images, and diffusion weighted images. The images are segmented and cropped to 512×512 as the Ground truth, and the bicubic downsampled or frequency domain downsampled images to 64×64 (8 times) or 128×128 (4 times) or 256×256 (2 times) are used as the input of the training set.

[0060] In this embodiment, the Adam optimizer with a momentum parameter of 0.9 is used to optimize the model parameters. The initial value of the learning rate is set to 0.0001, and the learning rate is adjusted every 100 epochs. Each input randomly crops out a 32×32 area, and the data is enhanced by random rotation and flipping.

[0061] The content of the comparative experiment is as follows:

[0062] 1. Explore the influence of network depth on network performance. Without multi-scale upsampling, only increase the number of cascade blocks and convolutional blocks. The results show that as the network deepens, the growth rate of the quality of the reconstructed image becomes smaller and smaller, while the parameter memory increases exponentially. When the number of cascade blocks is increased from 3 to 4, the PSNR only increases by 0.3db, while the parameter memory increases by 133%. Therefore, in subsequent experiments, the network depth is locked at 3 cascade blocks and 3 convolutional blocks.

[0063] 2. Explore the influence of the parameter-sharing attention mechanism and upsampling structure on network performance. The number of upsampling structures increases from the last convolutional block to 4, and the recorded results are as follows:

[0064]

[0065] As the number of upsampling features increases, the performance of the network improves. The difference in performance between parameter sharing and non-sharing is not significant, and parameter sharing makes the model 55% lighter in terms of parameter memory.

[0066] 3. Compare with existing advanced methods (SRGAN, CARN, SAN, RFDN, LESRCNN) as follows:

[0067]

[0068] The method proposed by the present invention has achieved obvious advantages in terms of parameter memory, running time, and image reconstruction quality. As Figure 3 shown, it shows the performance of different methods in the super-resolution reconstruction tasks of 2-fold (the first row) and 4-fold (the second row) MR images.

[0069] On the other hand, the present invention also discloses a magnetic resonance image super-resolution reconstruction system, including an image receiving module, a feature extraction module, a sampling module, an information integration module, and an image generation module. As Figure 4 shown:

[0070] The image receiving module receives 3D magnetic resonance images and segments them into 2D images, which are sent to the feature extraction module as single channels.

[0071] The feature extraction module receives the 2D images for deep feature extraction, outputs feature maps, and sends them to the sampling module.

[0072] The sampling module receives the feature maps, uses the spatial attention mechanism to weight their spatial positions, and performs multi-scale upsampling to obtain multiple groups of super-resolution feature maps, which are sent to the information integration module.

[0073] The information integration module performs integrated processing on the super-resolution feature maps and outputs a single-channel ultra-clear image to the image generation module.

[0074] The image generation module outputs the ultra-clear image.

[0075] On yet another aspect, the present invention also provides a magnetic resonance image super-resolution reconstruction device, mainly including a processor, a memory, and a bus. The memory stores instructions that can be read by the processor, and the processor is used to call the instructions in the memory to execute a magnetic resonance image super-resolution reconstruction method. The bus connects the functional components to transmit information.

[0076] In a preferred embodiment, the processor configuration information can be as follows:

[0077] The CPU is 10700k with a frequency of 3.8 MHz; the running memory is 64 GB with a frequency of 3200 MHz; the GPU is NVIDIA GeForce GTX 3080 (10 GB); the hard disk is Samsung 970 EVO Plus NVMe M.2 (1 TB).

[0078] In another implementation manner of this solution, it can be implemented by means of a device, and the device may include corresponding modules for executing each or several steps in the above-mentioned various implementation manners. Therefore, each step or several steps in the above-mentioned various implementation manners can be executed by the corresponding modules, and the electronic device may include one or more of these modules. The module may be one or more hardware modules specifically configured to execute the corresponding steps, or implemented by a processor configured to execute the corresponding steps, or stored in a computer-readable medium for implementation by the processor, or implemented through a certain combination.

[0079] The device can be implemented by using a bus architecture. The bus architecture may include any number of interconnected buses and bridges, depending on the specific application of the hardware and the overall design constraints. The bus connects various circuits including one or more processors, memories, and / or hardware modules together. The bus can also connect various other circuits such as peripheral devices, voltage regulators, power management circuits, external antennas, etc.

[0080] The bus can be an Industry Standard Architecture (ISA) bus, a Peripheral Component Interconnect (PCI) bus, an Extended Industry Standard Component (EISA) bus, etc. The bus can be divided into an address bus, a data bus, a control bus, etc. For the sake of representation, only one connecting line is shown in this figure, but it does not mean that there is only one bus or one type of bus.

[0081] Any process or method description shown in the flowchart or described in other ways herein can be understood as representing a module, segment, or part of code including one or more executable instructions for implementing a specific logical function or process, and the scope of the preferred implementation manner of this solution includes additional implementations, where the functions can be executed in a manner that is not in the order shown or discussed, including in a substantially simultaneous manner or in the reverse order according to the functions involved, which should be understood by those skilled in the art to which the implementation manner of this solution belongs. The processor executes the various methods and processes described above. For example, the method implementation manner in this solution can be implemented as a software program, which is tangibly included in a machine-readable medium, such as a memory. In some implementation manners, part or all of the software program can be loaded and / or installed via the memory and / or the communication interface. When the software program is loaded into the memory and executed by the processor, one or more steps in the method described above can be executed. Alternatively, in other implementation manners, the processor can be configured to execute one of the above methods in any other appropriate manner (for example, by means of firmware).

[0082] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that they can still modify the technical solutions described in the foregoing embodiments, or perform equivalent replacements on some or all of the technical features; and these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for super-resolution reconstruction of magnetic resonance images, characterized in that, It includes the following steps: (1) Convert the 3D magnetic resonance image into a single 2D magnetic resonance image as the single-channel input of the super-resolution network; (2) Convert the single-channel input into a multi-channel feature map, and use multiple cascaded blocks and convolutional blocks alternately arranged to progressively extract the depth feature map and , where , , represents the operation of the -th cascaded block, and represents the operation of the -th convolutional block; specifically, the cascaded block includes: further extracting the input features by combining 3 residual blocks and 3 3×3 convolutional blocks alternately, and connecting the features output by the residual blocks and the inputs of the 3×3 convolutional blocks in a cascaded manner; (3) Assign the same weight to the same spatial position in different depth feature maps and use a sub-pixel convolutional block with partial or full parameter sharing to upsample the depth feature maps at multiple scales to obtain multiple groups of super-resolution feature maps ; The specific process includes: using a spatial attention mechanism to weight the spatial positions of the depth feature maps output by each convolutional block, and using max pooling and average pooling to process the features at the same spatial position of each channel in, fuse the depth feature maps and then perform convolutional processing, and then convert the obtained features into weights between 0 and 1 by a function; The convolutional block in the spatial attention mechanism uses a 7x7 convolutional block; The specific expression is: ； Among them, represents different depth feature maps after being processed by the spatial attention mechanism; (4) The super-resolution feature maps at different depths are fused and processed using an information integration module to obtain a super-resolution reconstructed magnetic resonance image , where , is the operation of the information integration module, is the total number of convolutional blocks; the information integration module specifically includes: a feature integration structure composed of 3, 4, or 5 groups of RELU activation layers and 3×3 convolutional blocks alternatingly, where the last convolutional block outputs a single-channel magnetic resonance image.

2. The method for super-resolution reconstruction of magnetic resonance images according to claim 1, characterized in that, As described in step (4) is limited to the range of .

3. The method for super-resolution reconstruction of magnetic resonance images according to claim 1, wherein, The multi-channel feature map described in step (2) uses 64 channels.

4. A method for super-resolution reconstruction of magnetic resonance images according to any one of claims 1 to 2, characterized in that, During network training, the 3D magnetic resonance image is downsampled by bicubic interpolation by 2, 4, or 8 times as the input.

5. A method for super-resolution reconstruction of magnetic resonance images according to any one of claims 1 to 2, characterized in that, The sub-pixel convolutional block used in step (3) is controlled to perform upsampling by 2, 4, or 8 times as required.

6. A magnetic resonance image super-resolution reconstruction system using the method described in any one of claims 1 to 2, characterized in that, It includes an image receiving module, a feature extraction module, a sampling module, an information integration module, and an image generation module: The image receiving module receives the 3D magnetic resonance image and segments it into 2D magnetic resonance images to be delivered to the feature extraction module as a single channel; The feature extraction module receives the 2D magnetic resonance image for deep feature extraction, outputs a deep feature map, and delivers it to the sampling module; The sampling module receives the deep feature map, weights its spatial position using the spatial attention mechanism, and performs multi-scale upsampling to obtain multiple groups of super-resolution feature maps and send them to the information integration module; The information integration module performs integrated processing on the super-resolution feature maps and outputs a single-channel ultra-clear image to the image generation module; The image generation module outputs the ultra-clear image.

7. A magnetic resonance image super-resolution reconstruction device, characterized in that, It includes a processor, a memory, and a bus. The memory stores instructions that can be read by the processor; the processor is used to call the instructions in the memory to execute a method for super-resolution reconstruction of magnetic resonance images as described in any one of claims 1 to 2; the bus connects the processor and the memory and is used to transmit information.

Images